What is the NIST AI Risk Management Framework?
The NIST AI Risk Management Framework (AI RMF) is a voluntary guide from the US National Institute of Standards and Technology for managing the risks of AI systems. It was published in 2023 after public consultation with industry, academia, and government. It certifies nothing and no regulator enforces it directly, but it has become the common US reference for what a reasonable AI risk program looks like, and procurement questionnaires increasingly ask whether a supplier follows it.
The framework has two parts. The first describes the characteristics of trustworthy AI: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. The second, called the core, organizes the work into four functions.
How the NIST AI Risk Management Framework works
Govern establishes the culture and structure: policies, roles, accountability, and the process by which the other three functions happen. It is the only function that runs across the whole organization rather than per system. Map sets the context for one system: its intended purpose, the people it affects, the data it uses, and the risks that follow from those facts. Measure tests the risks that Map identified, using metrics, evaluations, red teaming, and tracking over time. Manage decides what to do about the measured risks, assigns resources, and plans the response when something goes wrong.
Each function breaks into categories and subcategories with suggested actions, and NIST publishes a companion playbook with practical steps. A separate profile for generative AI adds actions specific to language and image models, such as handling fabricated content and harmful outputs.
The NIST AI Risk Management Framework in practice
A property and casualty insurer adopts the framework for a claims triage model. Under Govern, the chief risk officer names a model owner and writes a one-page policy on when a human adjuster must review a claim. Under Map, the team documents that the model only prioritizes claims for review and never denies one, which keeps it out of the highest risk tier. Under Measure, they test referral rates across age groups and regions before launch and every quarter after. Under Manage, a measured disparity above an agreed threshold pulls the model from production until it is explained.
What to watch for
- The framework tells you what to consider and leaves the procedures to you, so expect to write your own.
- Use it to structure evidence for ISO/IEC 42001 certification or EU AI Act compliance, since the functions map cleanly onto both.
- A vendor that claims to be “NIST compliant” is describing alignment with a voluntary guide, so ask what it actually did.
- Start with Govern and Map, because teams that jump straight to metrics usually measure the wrong things.